4/1/2013. Environmental epidemiology: challenges. calculation of RR in cohort study. Research designs & bias in environmental epidemiology

Transcription

1 Research designs & bias in environmental epidemiology 1. challenges 2. designs traditional non-traditional 3. biases Sverre Vedal Environmental epidemiology: challenges 1. long latency 2. exposure measurement error 3. rare diseases 4. low-level exposure 5. small effect size notion of more probable than not Morgenstern & Thomas, EHP 1993 calculation of RR in cohort study Age (years) Death rate (deaths/ person-yrs) exposed subjects /200 5 unexposed subjects /250 exposed person-yrs = = 200 unexposed person-yrs = = 250 RR = 5/200 = /250 1

2 Environmental epidemiology: designs A problem: A public health worker in Turkey has been impressed by the large number of cases of mesothelioma in the region. How can this impression be moved to a more intellectually (scientifically?) rigorous level? types of study designs in environmental epidemiology traditional case series ecological cross-sectional case-control cohort non-traditional time series & case-crossover panel quasi-experimental (ie, natural experiment) gene-environment interaction intervention? Advantages of case series reports May identify new occupational or environmental hazard e.g., (besides asbestos and mesothelioma), vinyl chloride and angiosarcoma of the liver Can lead to intervention 2

3 Limitations of case series reports Only anecdotal information the plural of anecdote is not data May be a spurious cluster (Nassim Nicholas Taleb s Fooled by Randomness ) Ecological study design Approach correlate disease rates with geographical distribution of exposure Example compare rates of mesothelioma of Turkish towns with environmental asbestos to those of other towns Hasanoglu HC, et al. Int Arch Occup Environ Health 2005 Hasanoglu HC, et al. Int Arch Occup Environ Health

4 Hasanoglu HC, et al. Int Arch Occup Environ Health 2005 Ecological study design Hasanoglu HC, et al. Lung cancer and mesothelioma in towns with environmental exposure to asbestos in Eastern Anatolia. Int Arch Occup Environ Health areas Arguvan, Hekimhan rest of Malatya province mesothelioma cases (in 5 yrs) population incidence rate/100,00/yr 7 52, , Ecological study design Hasanoglu HC, et al. Lung cancer and mesothelioma in towns with environmental exposure to asbestos in Eastern Anatolia. Int Arch Occup Environ Health areas Arguvan, Hekimhan rest of Malatya province lung cancer cases (in 5 yrs) population incidence rate/100,00/yr 29 52, , is there a problem? 4

5 Ecological study design Main limitations no information on other risk factors for the disease (e.g., for lung cancer = age, smoking, occupational exposures) potential misclassification of people s actual exposures Cross-sectional design Approach: Describe/compare prevalence of some feature (e.g., disease/symptoms or physiological/imaging abnormality) typically by level of exposure A survey or slice in time without regard to exposure or health outcome Example: Compare chest x-ray abnormalities in Turkish towns with environmental asbestos compared to other towns Cross-sectional design Coplu L, et al. An epidemiological study in an Anatolian village in Turkey environmentally exposed to tremolite asbestos. J Environ Pathol Toxicol Oncol Approach Questionnaires and chest x-rays for all > 20 yrs old in village of Kureysler. Findings 18% had pleural plaques and/or calcification c/w asbestos exposure. 5

6 Cross-sectional design Coplu L, et al. An epidemiological study in an Anatolian village in Turkey environmentally exposed to tremolite asbestos. J Environ Pathol Toxicol Oncol Interpretation This is an example of a descriptive cross-sectional study in that it describes features of a population, but has no comparison group(s). More valuable if study included comparisons* with other communities. *[to give the authors credit, they state that chest x-ray screening in other communities detected no pleural abnormalities, but no details are provided.] Advantages of cross-sectional studies Good for less adverse outcomes (symptoms, physiological measures, imaging) Statistical power often good Direct contact with population sample (e.g., workers) permits additional individual-level data to be collected on: other risk factors of the outcome modifying factors [use of PPE, etc.] Limitations of cross-sectional studies Very susceptible to selection bias (e.g., migration between jobs/regions influenced by health status). Access only survivor population (e.g., workers most affected may have quit = healthy worker effect ). Whether exposure precedes the outcome (temporality) may be unclear. 6

7 types of study designs in environmental epidemiology traditional case series ecological cross-sectional case-control cohort non-traditional time series & case-crossover panel quasi-experimental (ie, natural experiment) gene-environment interaction intervention? Time Series Studies Compare day-to-day changes in exposure with total event numbers in an area Group-level design: Total event counts are the outcome and a representative measure of exposure is the population exposure a short-term exposure effect design the exposure Goldberg M et al. Am J Epidemiol. 2001;154:

8 the endpoint What would happen if you looked at the correlation between the two (ie, daily ozone and death counts?) Goldberg M et al. Am J Epidemiol. 2001;154: removing long-term temporal trends How is the Analysis Done? Aim: To relate daily event counts to daily exposure concentrations, adjusting for season and weather confounders Disease model for daily event counts Y t : log(e(y t )) = f 1 (time) + f 2 (weather) + DOW t )+ β(x t ) β is the parameter of interest for exposure x t f 1, f 2 are smooth functions Nonparametric smooth GAM model Parametric smooth GLM model 8

9 NMMAPS: the essential (multi-city) air pollution time series study: short-term PM 10 mortality associations in 88 US cities Dominici, Am J Epidemiol 2002 Time Series Study Designs Strengths Data are typically easy to obtain Individual-level factors are controlled by design Generally good statistical power Challenges Analytically complex Exposure measurement error that differs for different exposures can complicate interpretation Case-Crossover Studies Concept: The perfect control would be the case under a different exposure scenario (remember time machines) For transient exposures and outcomes with abrupt onset, each subject used as their own control Harvey Checkoway reviewed on March 27 9

10 Referent Time (control) Selection Critical in avoiding bias Might also want to match for temporal trends (day of week, month, season, etc.) Need to have rough understanding of timing so that select referent window outside of induction time or wash-out time Panel studies Longitudinal study where subjects (typically wellcharacterized) are followed (and measured) repeatedly over time A variant of the repeated measures design except many more measurements than is typical of repeated measures Example: n=188 children ages 6-13 followed for 18 months with daily measurement of symptoms and peak flow Vedal S et al. Acute Effects of Ambient Inhalable Particles in Asthmatic and Nonasthmatic Children. Am J Respir Crit Care Med 1998;157: Panel studies: example Vedal S et al. Am J Respir Crit Care Med 1998; 157:

11 Panel studies Advantages: 1. each subject their own control (so, no need to control for individual characteristics potential confounders) 2. can get measured endpoints (and lots of them) Disadvantages: 1. requires lots of resources to recruit and follow-up 2. analytically complex (eg, time trends need to be controlled) Quasi-experimental design (natural experiment) Takes advantage of an unnatural intervention that is not randomized. Examples: 1. pre-post study of water fluoridation comparing rate of dental caries in neighboring communities (Morgenstern & Thomas, 1993) 2. steel-mill strike reducing community air pollution levels (Utah Valley) A steel mill strike in the Utah Valley: a natural experiment on PM and hospitalizations Steel strike Utah Valley hospitalizations Pope CA, et al Pope CA. Arch Environ Health 1991; 46:

12 Quasi-experimental design (natural experiment) Advantages: 1. controls for many potential confounders 2. often easy to carry out Disadvantages: 1. since not really experimental, confounding is still a possibility Environmental epi intervention design Example: Romieu I et al. Antioxidant supplementation and lung functions among children with asthma exposed to high levels of air pollutants. Am J Respir Crit Care Med 2002; 166: Randomly assigned Vit E and Vit C to 158 asthmatic children in Mexico City to assess differential response to air pollution Finding: lung function declined in association with increases in ozone and PM in the placebo group but not the treatment group Case-control design Approach Compare past exposure of persons with disease (cases) with exposure of persons free of disease (controls). Example Compare history of residence in Turkish towns with high environmental asbestos in those with and without mesothelioma. 12

13 Design of case-control study time direction of research exposed not exposed exposed not exposed cases (people with the disease) controls (people without the disease) Population Exposure measurement Start Advantages of case-control design relatively quick and inexpensive good for rare diseases or diseases with a long latency period can examine more than one exposure simultaneously Limitations of case-control design inefficient if exposure is rare direct computation of incidence rates is not possible possible to confuse the temporal relationship between the exposure and the outcome particularly prone to bias (especially selection and recall [ information ] bias) 13

14 Cohort design Approach Compare rate of new disease in a well-defined population cohort according to degree of exposure. Example Compare mesothelioma rates in cohort of people from Turkish towns with varying degrees of environmental asbestos. Design of cohort study time direction of research disease Population COHORT subjects without the disease exposed not exposed no disease disease no disease Start Exposure measurement Follow-up heavy arsenic lung cancer death yr1 yr2 yr3 yr4 yr5 other death light arsenic lost yr1 yr2 yr3 yr4 yr5 incidence rate (IR) = (# cases)/(person-yrs) IR (exposed) = 2/25 = IR (unexposed) = 1/27 = IR ratio (IRR) = 2.16 = RR 14

15 Advantages of cohort studies good when rare exposure when prospective, less chance of exposure information bias time relationship clear (especially if prospective) can examine more than one outcome directly measures incidence of outcome in exposed and unexposed (if internal comparison) Limitations of cohort studies not good for rare diseases relatively costly and time-consuming (especially if prospective) if retrospective, often need records and the availability/quality of information may be a problem loss to follow-up may cause bias 15